What I've Built

Pitched executive leadership on a digital avatar system and scoped the MVP to a single quarter: small enough to move fast, big enough to validate. The MVP generated extraordinary community excitement: Reddit users had been craving a way to give themselves a "face" on a platform that had never offered one.

Scaled from MVP into a full identity platform: a layered SVG and PNG compositing system with constraint management, cross-platform rendering optimization, and a content management system that eventually held thousands of design elements. The customization engine was architecturally informed by prior experience building a garment customization system at Bonafides: layered image compositing with constraint management, translated from apparel into digital identity.

Built an entertainment partnership program tied to trending cultural moments, creating a re-engagement flywheel every time users returned to update their identity. Premium Avatar content behind a paywall became a meaningful retention contributor for existing Premium subscribers. Restructured the team into two pods (Core Avatar Experience and Distribution & Engagement) to separate product depth from growth velocity as the org scaled.

Led the Conversion Experience (ConvEx) team, owning the full top-of-funnel: SEO experience, authentication, and early-session retention. The core insight was that Reddit's Post Detail Page (PDP), which SEO users landed on, did nothing to help first-time visitors understand what Reddit was or why they should care. It assumed familiarity it hadn't earned.

Redesigned the PDP to meet first-time visitors where they were, progressively introducing platform concepts as users demonstrated readiness through their behavior rather than assuming familiarity the product hadn't earned.

This balancing act was extraordinarily sensitive. At SEO scale, even marginal changes in engagement metrics on landing pages had outsized downstream effects on acquisition, so every experiment required precision measurement, careful holdout design, and rigorous second-order impact analysis. A core finding that shaped the approach: the first session was rarely the right moment to push authentication, and users converted meaningfully better when prompted later in their lifecycle, once they had experienced enough of the product to understand its value.

Led AI-native product work inside Reddit's growth organization, building LLM-powered systems that improved new-user experiences and expanded content operations. The work spanned personalization, content pipelines, and internal tooling, informed by a strong point of view that prompt engineering should be a product surface rather than an engineering concern, with non-engineers given direct control over model behavior.

Ran the full product lifecycle across these systems: designing the human-in-the-loop touchpoints, building evaluation rubrics and review layers, and iterating quickly with a very small team. Outcomes contributed to measurable improvements in retention and unit economics for paid acquisition audiences.

Conceived and designed Sonia from scratch: a fully realized AI character with opinions, taste, and initiative, who maintains a persistent model of the user's life — their people, rhythms, goals, deferred intentions, and evolving preferences — and acts on it without being asked. The product philosophy was radical: the character is the product. Not a feature layer on top of an LLM, but a designed personality with four distinct conversational registers (friend, coach, therapist, discipliner) that unlock progressively as trust accumulates over weeks and months.

Defined the complete product system across three interlocking loops. The onboarding loop: an AI-driven conversation designed to feel like a first meeting with someone perceptive, engineered around a deliberate trust-building moment where Sonia references something mentioned earlier in a way that feels genuinely human. The weekly engagement loop: Sonia initiates a recurring conversation about what happened, what's coming, and what patterns she's noticed, then generates fully formed recommendations — specific venues, times, and reasons — delivered through a theatrical wait sequence designed to make the user feel Sonia "went somewhere" to find the answer. The living calendar: a three-state planning system (purgatory, real, horizon) where Sonia's working ideas float as cloud elements until the user commits, with every approval, deferral, and rejection feeding back into the life graph.

Built a working web prototype in Next.js with streaming Claude integration, persistent SQLite context model, real venue sourcing via Google Places and Bandsintown, proactive recommendation delivery via cron triggers, and a full transparency page where users can see and edit everything Sonia knows about them. Designed the complete character identity — visual language, animation states, voice principles — and built an agentic image generation pipeline using multimodal AI review loops to iteratively produce character assets at the quality bar the product demanded.

Co-founded Humtap and led product from seed through Series A. The core product: users provided melody and rhythm prompts, and the composition and sound design engines interpreted those prompts to generate full songs in various styles, an early generative creation platform predating the short-form video AI creation wave by several years.

Authored the fundraising pitch, product narrative, and deal documentation that closed a $12M Series A led by Mubadala (UAE sovereign wealth fund), working directly with investment bankers and legal counsel on deal structure. Post-funding, hired a distributed engineering team across London and Krakow: MLEs, mobile and backend engineers, and music operations (composers, sound designers for training data and quality benchmarks). Built the core data infrastructure and measurement systems from scratch.

During this period, Humtap was selected by Apple for the WWDC 2018 developer showcase. Spent several weeks at Apple's labs building a prototype integrating Humtap's generative AI composition engine with Apple's platform, validating that the product was at the frontier of what was possible in consumer AI at the time.

Designed the complete monetization system from the ground up. Built the Coins virtual currency infrastructure: coin pack pricing, purchase funnel, and the Award selection and granting UX distributed across feed and comment threads. Created the visual content library for Awards, a brand-iconic asset library still in use across Reddit today.

Designed the Free Awards mechanic as a deliberate try-before-you-buy funnel: users experienced the complete product loop (grant award, see the reaction, receive a thank-you) without paying anything. This built behavioral muscle memory before the purchase prompt ever appeared, meaningfully expanding the awarding population before any conversion pressure.

Built Reddit Premium from concept to MVP to scale: an annual subscription bundling the ad-free experience with a monthly Coin allocation sized to match the natural spending behavior of consistent purchasers. Distribution strategy was the primary growth lever. In-feed upsells targeted at high-usage Redditors and existing coin purchasers converted at multiples of cold promotion, reaching users at the exact moment they were experiencing platform value.